Chunking playground: how RAG splits your documents.
RAG does not search whole documents but sections: so-called chunks. Set chunk size and overlap and see at once how your text is split. That shows why chunking has a big influence on answer quality.
Runs in your browser · nothing is stored
Paste your text and see the chunks.
Chunks (overlap highlighted)
What is chunking in RAG?
When ingesting, a RAG platform splits every document into smaller sections and stores them in the search index. For each question it finds the best-matching sections and gives them to the language model. The size of these sections decides whether the model gets focused, complete context, or ill-fitting fragments.
Chunk size, overlap, word boundaries.
Chunk size
Small chunks match precisely but lose context. Large chunks keep context but contain more that does not fit.
Overlap
Part of each chunk repeats the end of the previous one. That way no sentence breaks at the boundary.
Word boundaries
Cuts at spaces instead of mid-word make chunks more readable and easier to find.
Good starting values for your documents.
| Document type | Chunk size | Overlap |
|---|---|---|
| FAQs, short answers | 300–500 characters | 10% |
| Policies, handbooks | 600–1,000 characters | 10–20% |
| Contracts, technical texts | 800–1,500 characters | 15–20% |
Rule of thumb: one token is about four characters. Test the values with real questions from your team.
Chunking you do not have to build yourself.
RAGSuite splits documents automatically during ingestion and keeps the index current when sources change. With the Quality Loop you see which questions are answered poorly and can improve them in a targeted way.
Questions about chunking.
Why does chunking matter so much?
Because RAG searches chunks, not whole documents. If chunks are too small, the model lacks context. If they are too large, they contain a lot that does not fit and push out better passages. Good chunking is one of the biggest levers for the answer quality of a RAG platform.
Characters or tokens: what does the tool use?
The tool splits by characters, because that can be shown exactly and transparently. As a rule of thumb one token is about four characters, so a 600-character chunk is roughly 150 tokens. German texts often have slightly more tokens. For a closer count, use the tokenizer.
What overlap makes sense?
Usually 10 to 20 percent of the chunk size. Overlap makes sure sentences at the boundary appear in full in both chunks. Too much overlap creates duplicate content and higher embedding costs. In the playground the overlapping part is highlighted, so you can compare the effect directly.
Is there a perfect chunk size?
No. It depends on your documents and how your team asks. Short FAQs need smaller chunks than contracts. Start with 600 to 1,000 characters and 10 to 20 percent overlap and check the results with real questions. Then adjust in a targeted way.
Is my document uploaded?
No. The text stays in your browser and the splitting happens locally. So you can test internal documents too, without anything leaving your computer. Close the page and the text is gone. This applies to all developer tools on ragsuite.de, without exception.
Good chunking is built in.
RAGSuite splits, indexes and updates your documents automatically. Install the Community Edition or try the RAGSuite Engine.